[Openmp-commits] [openmp] [OpenMP][NFC] Remove `DelayedBinDesc` (PR #74360)

[Johannes Doerfert via Openmp-commits]

jdoerfert wrote:

> > > > > This was the original issue #60119. I think HSA might've solved this as well in future versions.
> > > > 
> > > > 
> > > > This code was from the very beginning either unused or racy. I don't see a point in keeping it. If the issue exists, we need to store delayed lib initialization in a different place.
> > > 
> > > 
> > > I'll wait until @JonChesterfield chimes in, since if I recall this patch did indeed resolve the issue caused by infinitely recursive constructor calls.
> > 
> > 
> > The part I am deleting did never resolve an issue w/o race or UB. The critical part in the original patch was to move the loadRTLs call earlier. I included the reproducer of the original issue (#60119) which was not in our tests. It passes just fine on my machine w/o this code.
> 
> IIRC the bug was triggered by HSA's handling of that test case rather than the test itself. HSA was doing some evil binary trawling to search for whether or not it had to stand up debugging utilities or something if memory serves. This ended up causing the binary to re-load its constructors.

That might be true, but 1) the code in question was, and is, still UB or dead, and 2) the test should have been included.

https://github.com/llvm/llvm-project/pull/74360